MIST (Selective Trust): leaderboard
Selective-trust benchmark: 1,000 reasoning items presented in four matched conditions (clean, misleading, correct-context, irrelevant) to test whether a model uses helpful context while rejecting misleading signals; overall accuracy % across the conditions.
Metric: Overall Accuracy (%). Source: worldbench.github.io. Saturation forecast: Estimated already saturated. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 94.2 | #65 |
| 2 | GPT-5.5 | 94.2 | #33 |
| 3 | Claude Opus 4.8 | 93.7 | #40 |
| 4 | Claude Sonnet 4.6 | 92.8 | #96 |
| 5 | Claude Opus 4.7 | 91.1 | #52 |
| 6 | Claude Haiku 4.5 | 90.1 | #288 |
| 7 | Qwen 3 14B | 89.9 | #542 |
| 8 | Phi-4 (Reasoning) | 88.3 | #708 (Phi-4) |
| 9 | Qwen 3 8B | 88.1 | #679 |
| 10 | Qwen 2.5 14B | 85.5 | #716 |
| 11 | Seed 1.8 | 85.5 | #142 |
| 12 | Gemma 3 12B | 85.4 | #674 |
| 13 | Qwen 2.5 7B | 83.3 | #898 |
| 14 | Llama 3.1 8B | 76 | #1131 |
| 15 | Qwen 2.5 3B | 73.4 | #1177 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mist-selective-trust · How It Works · Data refreshed daily, snapshot 2026-11-06.